Research map / Video / Video generation
Veda: Scalable Video Diffusion via Distilled Sparse Attention
Shihao Han, Hao Yang, Xiaofeng Mei, Xinting Hu, Yi Jiang, Xiaojuan Qi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use oracle full-attention tile mask instead of average-pooling mask at fixed 90% sparsity (ablation) | sparse attention in video DiT · video generation quality | visual quality | Improved |
| add distilled tile scoring (Veda mask) over average pooling | sparse attention · tile recall vs full-attention oracle (Fig. 5) | tile recall: 34.2% -> 66.4% | Improved |
| use triplet (mean, max, min) tile statistics instead of average pooling (ablation) | Veda score estimator · estimator training loss | training loss: 0.965 -> 0.912 | Improved |
| add mean to max/min statistics (ablation) | Veda score estimator · estimator training loss | training loss: 0.982 -> 0.912 | Improved |
| use head-aware tiling instead of best static tiling [4,4,8] (ablation) | Veda · human evaluation, Waver-bench | Motion Quality / Overall Quality: +7.2% Motion Quality, +9.6% Overall Quality | Improved |
| stop gradient from estimator into backbone features (ablation) | Veda training · generation quality | generation quality | Improved |
| Veda 90% sparsity vs full attention | Waver-T2V-1B · human SBS, Waver-bench 1.0 | visual quality preference: 50% vs 49% | No clear effect |
| Veda vs VSA at equal 95% sparsity | Waver-T2V-1B · human SBS, Waver-bench 1.0 | visual quality preference: 76% vs 24% | Improved |
| raise Veda sparsity from 90% to 95% (ablation) | Wan2.1-T2V-1.3B · VBench | subject consistency: 0.887 -> 0.790 | Hurt |
| raise Veda sparsity from 90% to 95% (ablation) | Waver-T2V-1B · VBench | subject consistency: 0.940 -> 0.934 | Mixed |
| Oracle mask at 90% sparsity | Wan2.1-T2V-1.3B · VBench | aesthetic quality: 0.629 -> 0.325 | Hurt |
| sparse attention with Veda (95% sparsity) vs dense FA3 | Waver-T2V-12B · 720P 241 frames | layer latency: 1576.5ms -> 309.1ms | Improved |
Closest papers by meaning
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.93)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.91)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.90)
- Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · similarity 0.90)
- PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers (NeurIPS 2026 · similarity 0.90)
- VORTA: Efficient Video Diffusion via Routing Sparse Attention (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
